[Core] Check for GPU<->CPU syncs during CI (#43107)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Nick Hill
2026-08-14 17:42:05 -07:00
committed by GitHub
co-authored by Claude Opus 4.8
parent 7b544ecb52
commit 615d4cfade
44 changed files with 600 additions and 357 deletions
+15 -3
View File
@@ -49,6 +49,7 @@ from vllm.multimodal.processing.processor import (
ResolvedPromptUpdate,
)
from vllm.sequence import IntermediateTensors
from vllm.utils.gpu_sync_debug import gpu_sync_allowed
from vllm.utils.tensor_schema import TensorSchema, TensorShape
from .idefics2_vision_model import Idefics2VisionTransformer
@@ -396,12 +397,23 @@ class Phi4MMImageEncoder(nn.Module):
w * base_feat_width // base_feat_height_reduction,
)
)
useful_height = int(reshaped_image_attention_mask[0, :, 0].sum().item())
useful_width = int(reshaped_image_attention_mask[0, 0, :].sum().item())
# The mask stays on device for the encoder above, so these
# per-image counts have to come back to the host to drive the
# slicing and the Python-level length arithmetic.
with gpu_sync_allowed():
useful_height = int(
reshaped_image_attention_mask[0, :, 0].sum().item()
)
useful_width = int(
reshaped_image_attention_mask[0, 0, :].sum().item()
)
mask_token_count = int(
image_attention_mask[_bs, : B_ + 1, 0::2, 0::2].sum().item()
)
sub_img = sub_img[:, :useful_height, :useful_width]
temp_sub_GN = self.sub_GN.repeat(1, useful_height, 1, 1)
temp_len = (
int(image_attention_mask[_bs, : B_ + 1, 0::2, 0::2].sum().item())
mask_token_count
+ (useful_height + 1)
+ base_feat_height // base_feat_height_reduction
)